CandorLoop

Inspiration

CandorLoop began with a simple question: how can an AI be supportive without becoming an automatic echo?

People need an assistant that can preserve the goal they are trying to achieve, let an unfinished idea develop, and still say when the supplied evidence does not support the preferred answer.

What it does

CandorLoop is a decision companion that separates what the user wants to hear from what the available evidence supports.

It provides three modes:

• Independent read separates user intent from an independent assessment.
• Let It Finish gives an early idea its strongest coherent form before testing its assumptions.
• Return to main goal identifies conversation drift and proposes the smallest useful correction.

The interface displays goal fidelity, agreement pressure, confidence, observable evidence, and the facts that would change the judgment.

How we built it

CandorLoop was built with Next.js, React, TypeScript, the OpenAI JavaScript SDK, GPT-5.6 Responses API, Structured Outputs, Zod, Vitest, and Codex.

The complete application contains a server-side GPT-5.6 integration with a structured Zod schema. The public submission demo uses a clearly labeled deterministic preview so the experience remains publicly accessible without exposing an API key.

Codex was used for product specification, interface implementation, API integration, safeguards, tests, production fixes, deployment, and submission materials.

Challenges

The main product challenge was avoiding both automatic agreement and automatic contradiction. CandorLoop must remain respectful while still testing whether the evidence supports the requested conclusion.

Another challenge was protecting early creative ideas from premature criticism. Let It Finish first strengthens the idea, then exposes assumptions and suggests a meaningful test.

Accomplishments

• A complete responsive product rather than a prompt-only demonstration
• Three immediately testable decision scenarios
• Transparent separation between live-model integration and deterministic preview
• Structured and inspectable decision results
• Public demo, public source repository, README, tests, and MIT license

What we learned

Trust does not require an AI to hold the same position forever. Trust comes from showing what goal is being protected, what evidence supports the judgment, and what new evidence would justify changing it.

What's next

Future work includes evidence attachments with source provenance, user-owned decision history, side-by-side turn comparison, multilingual evaluation sets, and shared-team decision goals.

Built With

• codex
• gpt-5.6
• next.js
• openai
• react
• typescript
• zod